Accuracy of a deep learning–based model for treatment recommendation in adult patients with skeletal Class III malocclusion

Abstract

Introduction: Treatment planning for adult patients with skeletal Class III malocclusion remains challenging because of overlapping diagnostic criteria and subjective weighting of skeletal vs soft-tissue considerations. This retrospective study aimed to develop and evaluate the accuracy of a convolutional neural network (CNN)-based image classification in predicting treatment approach and supporting orthodontists in deciding between orthodontic camouflage and orthognathic surgery.

Methods: Using 1826 pretreatment images of 166 adult patients with skeletal Class III malocclusion (86 camouflage and 80 surgical), a hybrid model was developed that combines both deep learning and machine learning. These images included lateral cephalometric and panoramic radiographs and 9 intraoral and extraoral photographs. Of note, 11 CNN models processed each image type to generate binary predictions that were combined into an 11-dimensional vector and classified using 7 conventional machine learning algorithms.

Results: Support vector machine, multilayer perceptron, logistic regression, k-nearest neighbor, and naive Bayes showed no statistically significant difference compared with random forest (P >0.05). Decision tree exhibited statistically significant inferior performance compared with random forest (P <0.01). Significance analysis indicated that soft-tissue photographs had a higher correlation with treatment decisions than that of cephalometric radiographs, although clinical validity requires expert confirmation.

Conclusions: A CNN-based ensemble model demonstrated high diagnostic accuracy for predicting camouflage vs surgical treatment in adult patients with skeletal Class III malocclusion within a single-center dataset.